Talk by Chelsea Finn and Sergey Levine from UC Berkeley. Given to the Redwood Center for Theoretical Neuroscience at UC Berkeley.
Policy search methods based on reinforcement learning and optimal control can allow robots to automatically learn a wide range of tasks. However, practical applications of policy search tend to rely on hand-engineered components for perception, state estimation, and low-level control. In this talk, we will present methods for learning policies that map raw, low-level observations, consisting of camera images and joint angles, directly to the torques at the robot’s joints. To do so, we use guided policy search with deep spatial feature representations to efficiently learn policies with only tens of minutes of interaction time. We will show policies learned by a PR2 robot for a number of manipulation tasks which require close coordination between vision and control, including inserting a block into a shape-sorting cube, screwing on a bottle cap, and lifting a bag of rice into a bowl using a spatula.